Skip to main content
Claude Code is the strongest harness on Anthropic models, with skills, hooks, plugins, subagents, workflows, and MCP. paddock runs the real Claude Code CLI, unmodified, at the version it pins (currently 2.1.300), in each agent’s sandbox.

Models

Claude Code runs Anthropic models over the Anthropic Messages API:
  • Platform models: claude-opus, claude-sonnet, and claude-haiku.
  • Project models of an anthropic or anthropic-compatible provider, named provider/model. See Models.
Claude Code has two model settings, chosen independently from either source:

Capabilities

A plugin is installed with Claude Code’s own claude plugin commands and taken up by the running process without a restart; an MCP server with claude mcp. Changes apply between turns; see Capabilities.

Configuration

The harness runs non-interactively: there is no one to answer an approval prompt.

Tools

Claude Code’s tools map to paddock’s typed tools: Reasoning arrives as reasoning messages, and items paddock does not map, such as a context compaction boundary, as unknown messages with their raw JSON.

Harness status

Claude Code reports every section of the harness status. Its Context section is what /context shows: usage by category, counted with the model endpoint’s token counting, and the MCP tools, custom agents, memory files, and skills that take space.

Background work

Claude Code keeps background shells, subagents, and workflows running after a turn ends. When one finishes, Claude Code may start a turn on its own: paddock records it as a harness turn, which consumes no message. Stopping a session ends the current turn but spares background work. An agent does not sleep while a background agent or workflow runs.